What I learned from looking at 900 most popular open source AI tools

Chip Huyen published an analysis of the open source AI ecosystem, focused exclusively on the stack around foundation models, revisiting a similar analysis she did four years ago.

She searched GitHub for the keywords gpt, llm, and generative ai, finding 118,000 results for gpt alone, then limited results to repositories with at least 500 stars: 590 for llm, 531 for gpt, and 38 for generative ai. After also checking GitHub trending and social media, she found 896 repositories. Of those, 51 are tutorials or aggregated lists, including dair-ai/Prompt-Engineering-Guide and f/awesome-chatgpt-prompts; the analysis covers the remaining 845 software repositories.

She describes the AI stack as three layers: infrastructure, model development, and application development, plus separate categories for model repos and applications built on existing models. She plotted cumulative repository counts month over month and reports an explosion of new tooling in 2023 after the introduction of Stable Diffusion and ChatGPT, with the curve appearing to flatten in September 2023. She lists three potential reasons: the 500-star threshold takes time to reach, most low-hanging fruits have been picked, and people have realized it is hard to be competitive in generative AI.

The 845 repositories are hosted on 594 unique GitHub accounts. Twenty accounts with at least four repositories each host 195 of the repositories, or 23%, which have gained a total of 1,650,000 stars. Nineteen of those top 20 accounts are organizations, three of them belonging to Google: google-research, google, and tensorflow. The only individual account in that group is lucidrains. Over 20,000 developers have contributed almost a million contributions to the repositories; the 50 most active developers made over 100,000 commits, averaging over 2,000 each.

Huyen writes that China's AI ecosystem has diverged from the US and that her earlier impression that GitHub wasn't widely used in China is no longer true. Six of the top 20 accounts originated in China: THUDM, OpenGVLab, OpenBMB, InternLM, OpenMMLab, and QwenLM. She also reports that 158 of the 845 repositories (18.8%) gained no new stars in the previous 24 hours, and 37 (4.5%) gained none in the previous week, a pattern she and friends call the "hype curve."